Robust Tool Wear Estimation With Radial Basis Function Neural Networks

In this paper, a unified method for constructing dynamic models for tool wear from prior experiments is proposed. The model approximates flank and crater wear propagation and their effects on cutting force using radial basis function neural networks. Instead of assuming a structure for the wear model and identifying its parameters, only an approximate model is obtained in terms of radial basis functions. The appearance of parameters in a linear fashion motivates a recursive least squares training algorithm. This results in a model which is available as a monitoring tool for online application. Using the identified model, a state estimator is designed based on the upperbound covariance matrix. This filter includes the errors in modeling the wear process, and hence reduces filter divergence. Simulations using the neural network for different cutting conditions show good results. Addition of pseudo noise during state estimation is used to reflect inherent process variabilities. Estimation of wear under these conditions is also shown to be accurate. Simulations performed using experimental data similarly show good results. Finally, experimental implementation of the wear monitoring system reveals a reasonable ability of the proposed monitoring scheme to track flank wear.